Correct the recipe body and procedure pause axes in the module docs

[xmap]

Two documentation pages had drifted behind the conduct-path runtimes, in ways that would mislead someone planning work against them. Both were found while answering a design question about how CORA absorbs a change in how beamline staff run an experiment, which is exactly the question these pages are supposed to answer.

Recipe module page plus glossary

The RecipeStep union was documented as three arms when the code has five, and none of the four value-substitution sentinels appeared anywhere in the reference docs. The practical effect is a wrong conclusion: a reader would think values cannot flow between steps, when a capture step deposits a reading that a later setpoint consumes, and a compute step chains on an earlier step's artifact. What a recipe actually lacks is control flow, a much narrower and more interesting gap.

• Both missing arms documented (RecipeCaptureStep, RecipeComputeStep), with the fields each carries.
• New subsection keying the four sentinels on what distinguishes them: when each resolves (expansion versus execute time) and who produces the value. Only BindingRef is resolved by expand; the other three ride into the pinned step list and the determinism hash as opaque sentinels, which is what makes a steered or chained routine authorable as a recipe.
• The typed-not-textual substitution rule, the four wire keys, and the v1 "values bind, addresses do not" scope with its deferred triggers.
• The define-time structural checks, previously undocumented, including the one-sink guard that keeps the Dataset-of-record selection unambiguous. Both validators run in the handlers rather than the deciders, which is why their errors were missing from the slice error lists.
• The glossary gained Recipe and RecipeStep entries. It had neither, in a section titled "Recipe ladder".

Operation module page

An out-of-scope note claimed there is no pause-and-resume cycle today. Held has been real for a while: the status carries it, ProcedureHeld and ProcedureResumed are on the stream, hold_procedure and resume_procedure are slices, migration 20260621060000 widened the projection's status CHECK to admit it, and conduct_or_hold_procedure plus conduct_from_procedure exist only to serve it. The drift had spread past that one bullet:

• FSM diagram and transition table gain the pause pair, plus the asymmetry that abort and truncate accept Held while complete does not.
• Events table gains the five rows it was missing, not just the two pause events.
• The append-activities guard now admits that a paused conduct is not advancing.
• Aggregate state summary gains three fields.
• Projection DDL and its surrounding prose now match the migration, including that ProcedureResumed clears last_status_reason back to NULL.
• The Steering paragraph named reconduct_until_advised, a casualty of the re- prefix becoming a _from suffix. It was the only such survival anywhere in docs/.
• The title attribute hardcoded event and slice counts that had both gone stale. It now carries a qualitative claim like its sibling pages, and the event count moved into the body as a generated marker the build recomputes. Verified live by setting the source body to a wrong value and confirming the rendered page still read the computed one.

Verification

make docs-build (mkdocs --strict) builds clean, with no new anchor or link notices. pytest tests/architecture passes at 29733 passed, 591 skipped. Code was read as the authority throughout; no claim here is paraphrased from the prose being replaced.

🤖 Generated with Claude Code

[github-actions]

Coverage report

This PR does not seem to contain any modification to coverable code.